Embed Size (px)
Citation preview
Who am I?
• IOActive • Director of Penetration Testing • Pentest • Code review • Break stuff for fun and profit
2
agenda
• Intro • Observations • Problems • More observations • More problems summary • Solutions ?
3
What this talk is about
• Local security of linux on the desktop
4
Intro
• Used to use linux as my main desktop machine • Switched to windows about 7 years ago
• Mainly for work reasons
• Have used linux sporadically since then • However, not as a desktop OS
• Things seem to have changed somewhat
5
Intro
• Installed ubuntu during the holidays last year • Boots into a gui, looks like some mix of windows
and osx gui • It’s pretty clear they want to matter as a desktop
os • 7-10 years ago, it looked like, well, X • Today, it looks like something my grandmother can use
• Also looked at fedora, opensuse, knoppix • Also took a quick look at osx and opensolaris
(openindiana)
6
7
A decade ago
8
Now
Intro
• Initially had about a week or so of time to play around with this
• Made some interesting observations • Simple command line tools • Some code reading
• Found some clear problems • Maybe a solution or two
9
What was actually done (observations) • Started off with very simple commands to
enumerate some entrypoints • Wanted to see:
• shared memory (and it’s acl’s) (ipcs) • Udp/tcp/unix sockets exposed locally (netstat –pln) • Look at cron scripts • …
10
What was actually done (observations) • … • Wanted to see:
• Look for world writable files and directories • find / -perm -0666 -type f • find / -perm -0666 -type d
• Enumerate suid files • find / -perm +2000 -o -perm +4000 -type f
• Enumerate dbus system endpoints • dbus-send --system --type=method_call --print-reply --
dest=org.freedesktop.DBus /org/freedesktop/DBus org.freedesktop.DBus.ListNames
11
What was actually done (observations) • Expected this to be pretty boring and coming up
almost empty handed • Varying results for various distro’s and operating
systems • There seem to be some systemic issues across all
of them • Is no one doing trivial entrypoint analysis before
shipping ?
12
Overall finds (problems)
• Without disclosing details (bugs aren’t fixed) • world writeable shared memory • World writable scripts • Really really bloated suid binaries • misconfigurations • Over 60 finds in less than a week
• The goal of this talk isn’t any specific bug
13
More observations
• Dbus • Relatively new attack surface • X/Gnome/KDE specific • Ipc mechanism to pull information about the
system or the current session • Session is probably not that interesting • System could be!
14
More observations
• Dbus • Loads of new attack surface
• Configuration • Design (e.g. repurposing) • Implementation (e.g. buffer overflow)
• There seem to be piles and piles of these installed on default linux distro’s (40-60)
15
More observations
• Dbus system • Configure who can read / write to it
• Under what circumstances (root, console, group, default, …) • Where (what interface, …)
• /etc/dbus-1/* • Xml-alike file that specifies this
16
More observations
• Have been configuration bugs here in the past: • http://pkgs.fedoraproject.org/cgit/sectool.git/tree/sectoo
l-0.9.5-dbus.patch?id=aedb3ef7f7e5ab22d5438bfb7eee63489ccf3244;id2=4859832281f0e08c6fbe48fc252c4199a0e9e322
17
More observations
18
More observations
• Easy to make config mistakes • Similar to android intent permissions being set in
their AndroidManifest.xml file
19
More observations
• I wanted to look a bit closer at the suids • Asked readelf to give me a list of the library
dependencies (readelf –d) • All those libraries themselves are attack surface as
well • Some have just libc • Others depends on huge blobs of network parsers
(e.g. X).
20
More observations
21
ilja@ilja-VirtualBox:~$ readelf -d /usr/bin/kppp Dynamic section at offset 0x82ec8 contains 33 entries: Tag Type Name/Value 0x00000001 (NEEDED) Shared library: [libkde3support.so.4] 0x00000001 (NEEDED) Shared library: [libQt3Support.so.4] 0x00000001 (NEEDED) Shared library: [libkio.so.5] 0x00000001 (NEEDED) Shared library: [libkdeui.so.5] 0x00000001 (NEEDED) Shared library: [libkdecore.so.5] 0x00000001 (NEEDED) Shared library: [libQtCore.so.4] 0x00000001 (NEEDED) Shared library: [libQtDBus.so.4] 0x00000001 (NEEDED) Shared library: [libQtGui.so.4] 0x00000001 (NEEDED) Shared library: [libstdc++.so.6] 0x00000001 (NEEDED) Shared library: [libc.so.6] …
More observations
22
File: /usr/games/gnomine Dynamic section at offset 0x17ea4 contains 35 entries: Tag Type Name/Value 0x00000001 (NEEDED) Shared library: [libgtk-3.so.0] 0x00000001 (NEEDED) Shared library: [libgdk-3.so.0] 0x00000001 (NEEDED) Shared library: [libpangocairo-1.0.so.0] 0x00000001 (NEEDED) Shared library: [libpango-1.0.so.0] 0x00000001 (NEEDED) Shared library: [librsvg-2.so.2] 0x00000001 (NEEDED) Shared library: [libgio-2.0.so.0] 0x00000001 (NEEDED) Shared library: [libgdk_pixbuf-2.0.so.0] 0x00000001 (NEEDED) Shared library: [libgobject-2.0.so.0] 0x00000001 (NEEDED) Shared library: [libglib-2.0.so.0] 0x00000001 (NEEDED) Shared library: [libcairo.so.2] 0x00000001 (NEEDED) Shared library: [libpthread.so.0] 0x00000001 (NEEDED) Shared library: [libc.so.6] …
More observations
• I spend some time zooming in on those using X • X is a client/server protocol to be used for the gui in
most unices • Including most linux distributions
• It’s networked (can be tcp/ip, ipc, …) • Binary protocol
• suids in question are basically clients • You can make them connect to arbitrary X servers
using the DISPLAY variable • Who says these have to be well behaving X servers ?
23
X network architecture
24
More observations
• The X client libraries (Xlib) are clear attack surface • Spend about a day looking at the network parsing
code in Xlib • Things are really really bad
• Binary protocol parsers in C. • A lot of them written in the 80’s
• Server data appears to be trusted. Very little validation • > 60 trivial bugs
• It’s clear that code was not written with trust boundaries in mind at all. 25
More observations
26
Status XAllocColorCells( register Display *dpy, Colormap cmap, Bool contig, unsigned long *masks, /* LISTofCARD32 */ /* RETURN */ unsigned int nplanes, /* CARD16 */ unsigned long *pixels, /* LISTofCARD32 */ /* RETURN */ unsigned int ncolors) /* CARD16 */ { ... xAllocColorCellsReply rep; ... status = _XReply(dpy, (xReply *)&rep, 0, xFalse); if (status) { _XRead32 (dpy, (long *) pixels, 4L * (long) (rep.nPixels)); _XRead32 (dpy, (long *) masks, 4L * (long) (rep.nMasks)); } ... }
More observations
27
XTimeCoord *XGetMotionEvents(register Display *dpy, Window w, Time start, Time stop, int *nEvents) /* RETURN */ { xGetMotionEventsReply rep; ... if (!_XReply (dpy, (xReply *)&rep, 0, xFalse)) { … } ... if (! (tc = (XTimeCoord *) Xmalloc( (unsigned) (nbytes = (long) rep.nEvents * sizeof(XTimeCoord))))) { ... } ... for (i = rep.nEvents, tcptr = tc; i > 0; i--, tcptr++) { _XRead (dpy, (char *) &xtc, nbytes); tcptr->time = xtc.time; … } ...}
More observations
28
_XkbReadGetDeviceInfoReply( Display * dpy, xkbGetDeviceInfoReply * rep, XkbDeviceInfoPtr devi) { ... if (rep->nBtnsWanted>0) { act= &devi->btn_acts[rep->firstBtnWanted]; bzero((char *)act,(rep->nBtnsWanted*sizeof(XkbAction))); } ... int size; act= &devi->btn_acts[rep->firstBtnRtrn]; size= rep->nBtnsRtrn*SIZEOF(xkbActionWireDesc); if (!_XkbCopyFromReadBuffer(&buf,(char *)act,size)) goto BAILOUT; ... }
More observations
29
Private Bool _XimXGetReadData( Xim im, char *buf, int buf_len, int *ret_len, XEvent *event) { ... return_code = XGetWindowProperty(im->core.display, spec->lib_connect_wid, prop, 0L, (long)((length + 3)/ 4), True, AnyPropertyType, &type_ret, &format_ret, &nitems, &bytes_after_ret, &prop_ret); ... if (buf_len >= length) { (void)memcpy(buf, prop_ret, (int)nitems); ... } ... }
More observations
30
static char * ReadInFile(_Xconst char *filename) { register int fd, size; ... if ( (fd = _XOpenFile (filename, O_RDONLY)) == -1 ) return (char *)NULL; ... if ( (fstat(fd, &status_buffer)) == -1 ) { close (fd); return (char *)NULL; } else size = status_buffer.st_size; } if (!(filebuf = Xmalloc(size + 1))) { /* leave room for '\0' */ close(fd); return (char *)NULL; } size = read (fd, filebuf, size); ... }
More observations
31
static char* TransFileName(Xim im, char *name) { char *home = NULL, *lcCompose = NULL; char dir[XLC_BUFSIZE]; char *i = name, *ret, *j; int l = 0; while (*i) { if (*i == '%') { i++; switch (*i) { case 'H': home = getenv("HOME"); if (home) l += strlen(home); possible int overflow (long HOME and loads of %H's) break; case 'L': if (lcCompose == NULL) lcCompose = _XlcFileName(im->core.lcd, COMPOSE_FILE); if (lcCompose) l += strlen(lcCompose); possible integer overflow (long lcCompose and loads of %L's) break; case 'S': xlocaledir(dir, XLC_BUFSIZE); l += strlen(dir); possible integer overflow (long dir and loads of %S's) break;
More observations (cont)
32
j = ret = Xmalloc(l+1); integer overflow, alloc too short if any of the int overflows occured ... i = name; while (*i) { if (*i == '%') { i++; switch (*i) { case '%': *j++ = '%'; break; case 'H': if (home) { strcpy(j, home); buffer overflow if integer overflow occured j += strlen(home); } break; case 'L': if (lcCompose) { strcpy(j, lcCompose); buffer overflow if integer overflow occured j += strlen(lcCompose); } break; case 'S': strcpy(j, dir); buffer overflow if integer overflow occured j += strlen(dir); break;
More observations
33
XFixesCursorImage *XFixesGetCursorImage (Display *dpy) { ... xXFixesGetCursorImageAndNameReply rep; int npixels; ... if (!_XReply (dpy, (xReply *) &rep, 0, xFalse)) { ... } ... npixels = rep.width * rep.height; ushort * ushort ... rlength = (sizeof (XFixesCursorImage) + npixels * sizeof (unsigned long) + nbytes_name + 1); ... image = (XFixesCursorImage *) Xmalloc (rlength); ... image->x = rep.x; ... }
More observations
34
Bool DMXGetScreenAttributes(Display *dpy, int physical_screen, DMXScreenAttributes *attr) { ... xDMXGetScreenAttributesReply rep; ... if (!_XReply(dpy, (xReply *)&rep, (SIZEOF(xDMXGetScreenAttributesReply) - 32) >> 2, xFalse)) { UnlockDisplay(dpy); SyncHandle(); return False; } attr->displayName = Xmalloc(rep.displayNameLength + 1 + 4 /* for pad */); _XReadPad(dpy, attr->displayName, rep.displayNameLength); attr->displayName[rep.displayNameLength] = '\0'; ... }
More observations
35
int XGetDeviceButtonMapping( register Display *dpy, XDevice *device, unsigned char map[], unsigned int nmap) { int status = 0; unsigned char mapping[256]; /* known fixed size */ ... xGetDeviceButtonMappingReply rep; ... status = _XReply(dpy, (xReply *) & rep, 0, xFalse); if (status == 1) { nbytes = (long)rep.length << 2; _XRead(dpy, (char *)mapping, nbytes); ... } ... }
More observations
36
XEventClass * XGetDeviceDontPropagateList( register Display *dpy, Window window, int *count) { ... xGetDeviceDontPropagateListReply rep; ... if (!_XReply(dpy, (xReply *) & rep, 0, xFalse)) { ... } ... list = (XEventClass *) Xmalloc(rep.length * sizeof(XEventClass)); if (list) { ... for (i = 0; i < rep.length; i++) { _XRead(dpy, (char *)(&ec), sizeof(CARD32)); list[i] = (XEventClass) ec; } }
More observations
37
static int _XIPassiveGrabDevice(Display* dpy, int deviceid, int grabtype, int detail, Window grab_window, Cursor cursor, int grab_mode, int paired_device_mode, Bool owner_events, XIEventMask *mask, int num_modifiers, XIGrabModifiers *modifiers_inout) { ... if (!_XReply(dpy, (xReply *)&reply, 0, xFalse)) { ... } failed_mods = calloc(reply.num_modifiers, sizeof(xXIGrabModifierInfo)); ... _XRead(dpy, (char*)failed_mods, reply.num_modifiers * sizeof(xXIGrabModifierInfo)); for (i = 0; i < reply.num_modifiers; i++) { modifiers_inout[i].status = failed_mods[i].status; … }
More observations
38
XineramaScreenInfo * XineramaQueryScreens( Display *dpy, int *number ) { XExtDisplayInfo *info = find_display (dpy); xXineramaQueryScreensReply rep; .... if((scrnInfo = Xmalloc(sizeof(XineramaScreenInfo) * rep.number))) { ... for(i = 0; i < rep.number; i++) { _XRead(dpy, (char*)(&scratch), sz_XineramaScreenInfo); scrnInfo[i].screen_number = i; scrnInfo[i].x_org = scratch.x_org; scrnInfo[i].y_org = scratch.y_org; scrnInfo[i].width = scratch.width; scrnInfo[i].height = scratch.height; } ... }
More observations
39
XvImageFormatValues * XvListImageFormats ( Display *dpy, XvPortID port, int *num ){ ... xvListImageFormatsReply rep; ... if (_XReply(dpy, (xReply *)&rep, 0, xFalse) == 0) { ... } ... int size = (rep.num_formats * sizeof(XvImageFormatValues)); ... if((ret = Xmalloc(size))) { ... for(i = 0; i < rep.num_formats; i++) { ... } } ...
More observations
40
Bool XF86VidModeGetGammaRamp ( Display *dpy, int screen, int size, unsigned short *red, unsigned short *green, unsigned short *blue ) { ... xXF86VidModeGetGammaRampReply rep; ... if (!_XReply (dpy, (xReply *) &rep, 0, xFalse)) { ... } if(rep.size) { _XRead(dpy, (char*)red, rep.size << 1); _XRead(dpy, (char*)green, rep.size << 1); _XRead(dpy, (char*)blue, rep.size << 1); } ... }
More observations
41
Bool uniDRIOpenConnection(dpy, screen, hSAREA, busIdString) Display *dpy; int screen; drm_handle_t *hSAREA; char **busIdString; { ... xXF86DRIOpenConnectionReply rep; ... if (!(*busIdString = (char *)Xcalloc(rep.busIdStringLength + 1, 1))) { ... } _XReadPad(dpy, *busIdString, rep.busIdStringLength); ... }
More observations
42
static XcursorFileHeader * _XcursorReadFileHeader (XcursorFile *file) { ... if (!_XcursorReadUInt (file, &head.ntoc)) unsigned 32bit ntoc var read from file return NULL; ... fileHeader = _XcursorFileHeaderCreate (head.ntoc); passed on to allocate buffer ... for (n = 0; n < fileHeader->ntoc; n++) { if (!_XcursorReadUInt (file, &fileHeader->tocs[n].type)) break; if (!_XcursorReadUInt (file, &fileHeader->tocs[n].subtype)) break; if (!_XcursorReadUInt (file, &fileHeader->tocs[n].position)) break; } ... }
More observations (cont)
43
static XcursorFileHeader * _XcursorFileHeaderCreate (int ntoc) { XcursorFileHeader *fileHeader; if (ntoc > 0x10000) return NULL; fileHeader = malloc (sizeof (XcursorFileHeader) + ntoc * sizeof (XcursorFileToc)); if (!fileHeader) return NULL; fileHeader->magic = XCURSOR_MAGIC; fileHeader->header = XCURSOR_FILE_HEADER_LEN; fileHeader->version = XCURSOR_FILE_VERSION; fileHeader->ntoc = ntoc; fileHeader->tocs = (XcursorFileToc *) (fileHeader + 1); return fileHeader; }
More observations
44
static void ReqCleanup( Widget widget, XtPointer closure, XEvent *ev, Boolean *cont) { ... char *value; ... (void) XGetWindowProperty(event->display, XtWindow(widget), event->atom, 0L, 1000000, True, AnyPropertyType, &target, &format, &length, &bytesafter, (unsigned char **) &value); should check return value. XFree(value); ... }
More observations
• The X client libraries (Xlib) • All discovered X bugs are being fixed • The developer involved is actually quite good
• Amazing • Alan Coopersmith • Very deep understanding of X and the bugs involved • No pushback, no handholding • Worked tirelessly • 104 patches, with reviews and some tests in < 3 months
• And had some interesting comments
45
46
“I don't know how many setuid X clients still exist these days (is xterm still setuid on any platforms, or did they all get grantpt() or similar calls to avoid needing root?), but since we know there's more X clients than we can keep track of (especially once you get to home grown apps in various companies
they've been using for decades), we have to assume there still may be some. It would be good to put a reminder in the security advisory that best practice is to separate out the parts of an application that require elevated privileges from the
GUI to avoid such issues - GTK requires this, but not all toolkits do.”
More observations
More observations
47
“Shoot me now. And then shoot daniels for not freeing us from XKB yet. And then shoot anyone who volunteers to try to fix XKB, before it's too late for them too.”
“Here's my initial analysis of the first part of the Xlib set, until I got so tired my head started spinning trying to figure them out”
“Really, if your window shape is anywhere near 2^32 rectangles, what are you doing?”
“Yes, these [bugs] all seem possible, and far more feasible now than when this code was written, back when disk sizes were still measured in megabytes.”
More observations
• Comment on LWN from one of the people that introduced some of these bugs in the ‘80’s
48
More observations
• Debian’s turn around on these bugs was ridiculously fast!
• 104 patches to merge in • 2 week embargo • Full releases and advisory on day embargo expired
• No one else managed to do this
• I think Moritz Mühlenhoff deserves most of the credit for that one
49
More observations
• However, Raw X is rarely used nowadays • There’s stuff build on top • And they use raw X
• gtk+ • KDE (which uses QT which uses raw X) • Rare direct calls to Xlib code • other
50
More observations
• KDE / QT • Crappyness is about on par with Xlib • Trivial bugs!
51
More observations
• Several instances of this
52
Qkeymapper_x11.cpp void QKeyMapperPrivate::clearMappings() { ... uchar *data = 0; if (XGetWindowProperty(X11->display, RootWindow(X11->display, 0), ATOM(_XKB_RULES_NAMES), 0, 1024, false, XA_STRING, &type, &format, &nitems, &bytesAfter, &data) == Success && type == XA_STRING && format == 8 && nitems > 2) { ... char *names[5] = { 0, 0, 0, 0, 0 }; char *p = reinterpret_cast<char *>(data), *end = p + nitems; int i = 0; do { names[i++] = p; p += qstrlen(p) + 1; } while (p < end); ... }
More observations
• In QT init code (affects all QT applications)
53
Qapplication_x11.cpp void qt_init(QApplicationPrivate *priv, int, Display *display, Qt::HANDLE visual, Qt::HANDLE colormap) { ... } else if (arg == "-name") { if (++i < argc) appName = argv[i]; if it was previously new'ed, it isn't anymore. } ... } void qt_cleanup() { ... if (X11->foreignDisplay) { delete [] (char *)appName; could delete [] a pointer that isn't new'ed and possibly corrupt memory appName = 0; } ... }
More observations
• So I reported these bugs • They didn’t seem to think it was a security issues • Quote from X developer (previous slide) is dead on
• Suid remark
• QT does not seem to agree with this
54
More observations
• "KDE has precisely one setuid application, kcheckpass, for this reason. I suspect that someone running an suid Qt application would fall into a huge number of problems, the most obvious one being a malicious style that would allow them to trivially execute arbitrary code“
• Wait, did we just get a free Code exec bug from the QT security team ?
55
More observations
• I respond back, saying there are more KDE suid binaries, and specifically mention kppp, and question him on the styles thing
56
More observations
57
“> I am aware of this, regardless, this is library code, as such, chances are, there are suid applications out there that will use it. That would be a security hole in those applications rather than in Qt, there are many ways that people can abuse a library to create unsafe applications. > Do styles contain executable code ? Yes.”
58
“> also, does kppp no longer run suid ? kppp should not be installed setuid. Here's a quote from its FAQ: "There is no need for the setuid bit, if you know a bit of UNIX® systems administration. Simply create a modem group, add all users that you want to give access to the modem to that group and make the modem device read/writable for that group." I doubt any modern distro would install it suid, in fact most are extremely careful about what they allow to be suid and are actively working to minimise what is.”
More observations
More observations
• That kppp FAQ quote in incomplete, it goes on to say:
59
“… The KPPP team has lately done a lot of work to make KPPP setuid-safe. But it's up to you to decide if you install and how you install it.”
More observations
• In fact, distro’s do still have it suid. • E.g. Ubuntu • This is library code! They should not set policy for
the apps that use them. • They’re sitting on the fence, because it’s easy
• you don’t actually have to do anything
• Either defend it, and shut up • Or do a suid check and exit()
60
More observations
• None of those bugs are fixed • Got the ok from QT security team to disclose:
61
“> Ok, since you guys don't consider this a security issue, you're ok with me talking about this publicly? Yes, that's fine”
More observations
• So loaders have LD_PRELOAD • And has been made setuid safe years ago • KDE/QT
• QT_PLUGIN_PATH
• Gnome • GTK_MODULES
• Neither are setuid safe !
62
More observations
• The GTK+ people seem to be doing somewhat better.
• They do not allow suid GTK+ applications. • And clearly explain why on their webpage
63
More observations (http://www.gtk.org/setuid.html)
Talking Radios born from taco trucks
64
More observations
• This is beautiful, well though out and sane! • “Security of GTK+ requires the security of Xlib. The
GTK+ team is not prepared to make that guarantee”
• Or is it ?
65
More observations
66
Gtkmain.c /* This checks to see if the process is running suid or sgid * at the current time. If so, we don't allow GTK+ to be initialized. * This is meant to be a mild check - we only error out if we * can prove the programmer is doing something wrong, not if * they could be doing something wrong. For this reason, we * don't use issetugid() on BSD or prctl (PR_GET_DUMPABLE). */ static gboolean check_setugid (void)
More observations
• What does that mean ? • Suid binaries can use GTK+, BUT … • … they must acquire the privileged resources they
want first • And then drop privileges • After that it’s ok to use GTK+ • Want to have their cake and eat it too • Check should be stronger!
67
More observations
• games are a great example • They are suid
• Share a highscore database
• Once aquired, privs are dropped • Only thing an attacker would have access to is that db
• assuming a bug was found and exploited
• That db is considered trusted. • Any security bug in db parsing allows for further escalation
• Any user now playing any of those games gets pwned 68
More observations
• Spend a little bit of time looking at x display managers • There’s a lot of them • Uses xdmcp protocol (goes over udp) • Most have dependency on libxdmcp for this • Libxdmcp’s api’s quite easily lend themselves to
abuse • Leaves a lot of stuff uninitialized on failure
• This is being fixed
69
More observations
• LightDM • Used by ubuntu
• Has so called greeters that allow you to customize the gui
• Unpriv’ed greeters talk to LightDM • Using a pipe
• Parser for that pipe wasn’t great • Not that bad either
• Bugs are being fixed 70
More observations
71
static gboolean read_cb (GIOChannel *source, GIOCondition condition, gpointer data) { Greeter *greeter = data; gsize n_to_read, n_read, offset; ... n_to_read = HEADER_SIZE; if (greeter->priv->n_read >= HEADER_SIZE) { offset = int_length (); n_to_read += read_int (greeter, &offset); [2] } status = g_io_channel_read_chars (greeter->priv->from_greeter_channel, (gchar *) greeter->priv->read_buffer + greeter->priv->n_read, n_to_read - greeter->priv->n_read, [3] &n_read, &error); ... greeter->priv->n_read += n_read; if (greeter->priv->n_read != n_to_read) return TRUE; /* If have header, rerun for content */ if (greeter->priv->n_read == HEADER_SIZE) { gsize offset = int_length (); n_to_read = read_int (greeter, &offset); if (n_to_read > 0) { greeter->priv->read_buffer = secure_realloc (greeter, greeter->priv->read_buffer, HEADER_SIZE + n_to_read); [1] read_cb (source, condition, greeter); return TRUE; } } ... }
More observations
1. integer overflow when allocating buffer 2. Integer overflow (really small n_to_read) 3. Integer underflow, size to read becomes really large
72
More observations
• As mentioned earlier, libraries build on top of X use Xlib
• Apps will sometimes also call some X api’s to query certain things
• Using the XGetWindowProperty() api or any number of api’s build on top of it (e.g. XGetClassHint(), XGetRGBColormaps(), …)
• Looked at the use of Xlib api’s • This too wasn’t great
73
More observations
• By far the most common bug when using Xlib
• Check return values ! • XLib defense in depth fixes. Now guarantees NULL
init of arguments on failure. 74
void fn() { ... SomeFormat *sf; ... (void) XGetWindowProperty(dpy, w, property, 0L, 10000000, False, SomePropertyType, &type, &format, &length, &bytesafter, (unsigned char **) &sf); ... XFree((char*)sf); ... }
More observations
• Developers using Xlib don’t seem to realize that most of the api’s they use parse potentially untrusted network data
• _XReply
• _XRead32
• _XRead
• _XGetAsyncReply
• XGetWindowProperty
• XNextEvent
• XPeekEventXIfEvent
• XCheckIfEvent
• XPeekIfEvent
• XCheckTypedWindowEvent
• XSetErrorHandler
• XQueryFont
75
More observations
• derived from XGetWindowProperty: • XFetchName
• XGetIconName
• XGetSizeHints
• XGetWMHints
• XGetWMSizeHints
• XGetIconSizes
• XGetTransientForHint
• XGetClassHint
• XGetRGBColormaps
• XGetTextProperty
• XGetWMName
• XGetWMIconName
• XGetWMClientMachine
• XGetCommand
• XGetWMColormapWindows
• XGetWMProtocols 76
• XScreenResourceString • XFetchBuffer • XFetchBytes • XkbRF_GetNamesProp
More observations
• Conceptually there’s a couple of X suid apps around that you’ll see:
• Config tools (e.g. kppp) • Games (e.g. swell foop) • Screen locking utils (e.g. Xlock, Xlockmore, Xscreensaver, …)
• Virtually all of these apps do drop privileges
77
More observations
• the screenlocking utils • Only seem to capture your hashed pw entry (and
optionally root). • Getspnam()
• W00t. That’s not much of a resource • Or is it ?
78
More observations
• The linux shadow library is responsible for api’s for reading from and writing to the shadow file
• Shadow.h • The code uses FILE stream api’s to read and write
• Uses heap buffers internally, Can’t clear memory.
• Stores read data in local stack buffers • Doesn’t clear memory
• Basically leaks the entire shadow file onto the heap and the stack
79
More problems summary
• Xlib in suids is a bad idea • GTK+ kinda sorta still allowed in suids • Very common sloppy misuse of Xlib api’s • Linux shadow library handles sensitive data in a
sloppy manner
80
todo
• There’s some other things we wanted to look at but didn’t get around too
• Package managers • Clipboard
• Ok, we did look for about 5 seconds • It did look bad …. 81
todo
• Fedora package management (software)
82
todo
• Ubuntu apt-get
83
todo
• OpenSuse zypper
84
todo
• Lets hope they sign stuff • And check it
• And have signatures locally • Or fetch them securely from a remote host
• And don’t use md5 … • Even if you’ve got all this correct • Network protocol is unencrypted • Adds a lot of remote attack surface
85
todo
• Clipboard • ICCCM • It smells rotten • http://lists.slug.org.au/archives/slug-
chat/2001/July/msg00054.html • Srsly, go read it • No, really! • It’s all I know about ICCCM, but it speaks volumes
86
todo
87
“d00d, that document is devilspawn…. what sick evil twisted mind wrote this damn spec?”
“The ICCCM is the coding equivalent of the Medieval rack, except its advertised as some kind of X11 swingers party.”
“I've seen more elegant protocols in unlikely places. When blowflies fight over a pile of elephant shit, their pecking order is a more elegant protocol than ICCCM.”
“I. C. C. C. M. Inter- Client Communications Conventions Manual! Manual, like in "manual labour", like in "pain" Conventions, like in "not required, just do ALL OF IT or you SUCK!" Communications, like in "fucking overengineered carrier pigeons" Client, like in "see that guy with the limp, he was one of my ``clients''" Inter-, like in "Inter-nal bleeding“”
Solutions ?
• The shadow library thing is easy to fix • It’s not really a bug in the first place
• But exposes too much sensitive information to an already compromised suid program
• Drop all FILE stream usage. • Use open/read/write syscalls instead
• Clear all memory after use • Make sure memset() doesn’t get optimized out by compiler
88
Solutions ?
• Most suids on linux (and most unices) have been dropping privileges for a long time
• Nothing has changed since
• This isn’t good enough. • Those privileged resources include:
• read fd to /dev/kmem • read/write fd to /etc/resolv.conf • Full content of /etc/shadow • …
89
Solutions ?
• The suid processes still have their suid bit set in kernel • Attacker still needs some kind of bug + exploit • reduced what can be gained from uid 0 to those
resources • Is however still very significant
90
Solutions ?
• A model where priv dropping and priv separation is combined would make more sense
• Would add more defense in depth • Probably not that hard to implement for some suids
91
Solutions ?
• Here’s what it would looks like: • pipe • fork • client drops all privs • server gets resources • server drops privs • very small and well defined interface between client and
server
92
Solutions ?
• Client retains it’s suid bit • Pipe to server is protected from injection
• readelf –d on some of the suids • HUGE list of library dependencies!
• We don’t want that in service code. • fork() is out!
• If you fork, all that stuff is still in memory.
93
Solutions ?
• fork() is out, fork() + execve() is in. • Pass fd to excve’ed process. • Server has to be a very small piece of c code. • Only access to libc.
94
Solutions ?
• Actually, glibc (default on most distributions) is super bloated.
• > 100mb of source code (2.1.7) • memfrob() ? • strfry() ???
• Can you really trust that ? • Should not be used in server app • Instead use something like dietlibc, uClibc, klibc or
musl
95
Solutions ?
• One last piece of code bloat left • Dynamic loader. • Takes input though environment variables • Have been bugs in there in the past • Do you really want to trust it ? • Fix: static binaries
96
Conclusion?
• Guess there should be a conclusion • Run for the hills ? • Things could be better … • … by several orders of magnitude
• There’s really a lot of work to be done here
• most of that code is not written with a trust boundary in mind
97
98
Questions ?
